Applications of Statistical Learning and Stochastic Filtering for Damage Detection in Structural Systems

[摘要] Structural health monitoring (SHM) is necessary for online maintenance of infrastructure systems. It entails deployment of an array of sensors on a structure of interest for data acquisition, based on which, damage detection, quantification and localization may be performed (collectively known as condition assessment), followed by prognosis and estimation of remaining useful life that constitutes a comprehensive decision making framework. This dissertation specifically focuses on damage detection which traditionally follows a model-based approach. This implies construction of high-fidelity models resembling the actual system in its undamaged state followed by a quantitative or qualitative comparison of response from the model and the real infrastructure system. This sheds light on the changes in the system due to advent of damage. However, development of such high-fidelity models is extremely challenging and in many cases renders the damage detection problem computationally prohibitive. The alternative is using statistical learning and stochastic filtering algorithms that rely on construction of parametric models. Stochastic filtering methods are not necessarily model-free, however, they utilize data acquired from a real system for damage detection. Statistical learning approaches on the other hand either involve construction of either a meta-model or no model at all. The meta-models do not necessarily simulate system behavior in its full complexity, instead they are designed for the specific task of damage detection. If a sizable data is available from deployed sensors, these class of methods becomes far more suitable compared to the traditional model-based regimes. This dissertation proposes applications of statistical learning and stochastic filtering techniques for the task of damage detection, with an eventual goal of developing efficient SHM systems. The damage detection problems considered involve acoustic emission source localization in plates, active sensing of plates and pipe structures using Lamb waves and decoupling of effects of varying environmental conditions and damage on vibration testing data acquired from bridge structures. The proposed algorithms perform damage detection efficiently. Additionally, in the high frequency regime, the number of sensors necessary for damage detection is reduced.

[发布机构] Rice University

[效力级别] detection